The photo-sharing website Pinterest just rolled out a visual search tool that lets you zoom in on a specific object in a pinned image (or “Pin”) and discover visually similar objects, colors, patterns and more. For example, see a lamp that you like in a Pin of a living room? Tap the search tool in the corner of the Pin, drag the zoom tool over the lamp and scroll down for visually similar Pins.
A team of four engineers from the Pinterest Visual Discovery group and members of the Berkeley Vision and Learning Center worked in close collaboration to develop the core of the system in just a few months. The team used GPU-accelerated deep learning to teach their system how to recognize image features using a richly annotated dataset of billions of Pins curated by Pinterest users. The features can then be used to compute a similarity score between any two images and identify the best matches. For the past couple months, they’ve been experimenting with improving Related Pins using these visual signals, as detailed in their latest white paper, released this week.
As mentioned in the white paper, they are using the open-source Caffe deep learning framework to perform training and interference of their convolutional neural networks on multi-GPU machines.